Terren, >to the unembodied agent, it is not a concept at all, but merely a symbol with >no semantic context attached
It's an issue when trying to learn from NL only, but you can injects semantics (critical for grounding) when teaching through a formal_language[-based interface], get the thinking algorithms working and possibly focus on NL-to-formal_language conversions later. >To an unembodied agent, the concept of self is indistinguishable from any >other "concept" it works with. An AGI should be able to use tools (external/internal applications) and it can learn to view itself (or just some of its modules) as its tool(s). You can design an interface [possibly just for advanced users] for mapping learned concepts/actions to the interface of available tools. Just like it can learn how to use a command line calculator, it can learn how to use self as a tool. Then it can learn that an alias to use for "that tool" is "I"/"Me". By design, it can also clearly distinguish between using a particular tool "in theory" and "in practice". > All such an agent can do is perform operations on ungrounded symbols - at > best, the result of which can appear to be intelligent within some domain > (e.g., a chess program). You can ground when using semantic-supporting input formats. I don't see why would it have to be specific to a single domain. You can use very general data representation structures and fill it with data with many domains. You "just" have to get the KR right (unlike CYC). Easy to say, I know, but I don't see a good reason why it couldn't (in principle) work and I'm working on figuring that out. >> Even though this particular >> AGI never >> heard about any of those other tools being used for cutting >> bread (and >> is not self-aware in any sense), it still can (when asked >> for advice) >> make a reasonable suggestion to try the "T2" >> (because of the >> similarity) = coming up with a novel idea & >> demonstrating general >> intelligence. > > Sounds like magic to me. You're taking something that we humans can do and > sticking it in as a black box into a hugely simplified agent in a way that > imparts no understanding about how we do it. Maybe you left that part out > for brevity - care to elaborate? It must sound "like magic" when assuming the "no semantic context attached", but that doesn't have to be the case. With right teaching methods, the system gets semantics, can make models and can apply knowledge learned from scenario1 to scenario2 in unique ways. What does the "right teaching methods" mean? For example, when learning an "action concept" (e.g. "buy"), it needs to grasp [at least some] roles involved (e.g. "seller", "buyer", "goods", "price", ..) and how relationships between the role-players changes in relevant stages. You can design user friendly interface for teaching systems in meaningful ways so it can later think using queriable models and understand relationships [changes] between concepts etc... Sorry about the brevity (busy schedule). Regards, Jiri Jelinek PS: we might be slightly off-topic in this thread.. (?) ------------------------------------------- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com